Efficient Method for Trademark Image Retrieval: Leveraging Siamese and Triplet Networks with Examination-Informed Loss Adjustment

Image-based similar trademark retrieval is a time-consuming and labor-intensive task in the trademark examination process. This paper aims to support trademark examiners by training Deep Convolutional Neural Network (DCNN) models for effective Trademark Image Retrieval (TIR). To achieve this goal, we first develop a novel labeling method that automatically generates hundreds of thousands of labeled similar and dissimilar trademark image pairs using accompanying data fields such as citation lists, Vienna classification (VC) codes, and trademark ownership information. This approach eliminates the need for manual labeling and provides a large-scale dataset suitable for training deep learning models. We then train DCNN models based on Siamese and Triplet architectures, evaluating various feature extractors to determine the most effective configuration. Furthermore, we present an Adapted Contrastive Loss Function (ACLF) for the trademark retrieval task, specifically engineered to mitigate the influence of noisy labels found in automatically created datasets. Experimental results indicate that our proposed model (Efficient-Net_v21_Siamese) performs best at both True Negative Rate (TNR) threshold levels, TNR = 0.9 and TNR = 0.95, with respective True Positive Rates (TPRs) of 77.7% and 70.8% and accuracies of 83.9% and 80.4%. Additionally, when testing on the public trademark dataset METU_v2, our model achieves a normalized average rank (NAR) of 0.0169, outperforming the current state-of-the-art (SOTA) model. Based on these findings, we estimate that considering only approximately 10% of the returned trademarks would be sufficient, significantly reducing the review time. Therefore, the paper highlights the potential of utilizing national trademark data to enhance the accuracy and efficiency of trademark retrieval systems, ultimately supporting trademark examiners in their evaluation tasks.